Source-linked AI summary
Assistive Gym: A Physics Simulation Framework for Assistive Robotics
Zackory Erickson, Vamsee Gangaram, Ariel Kapusta, C. Karen Liu, Charles C. Kemp
TL;DR
Physical human-robot assistance research faces high costs and risks from physical interaction. Assistive Gym is an open-source simulation framework spanning six ADLs and four commercial robots. Modeling human movement through co-optimization increased task success by 30.4% on average.
Problem
Physical human-robot assistance research faces high costs and risks from physical interaction.
Source: §I · p. 1Method
Assistive Gym is an open-source physics simulation framework with six ADL environments that models human capabilities and preferences and supplies rewards for baseline policy training across four robots.
Source: §I · p. 1, E. Modeling Human Preferences · p. 4Results
Modeling human movement through co-optimization increased task success by 30.4% on average.
Source: §I · p. 1Takeaways & Limitations
The framework supports benchmarking, development of environments for multiple assistive tasks, and robot comparison, and is presented as promising for autonomous physical assistance.
Source: §VI · p. 6
Abstract
from arXiv · showhide
Autonomous robots have the potential to serve as versatile caregivers that improve quality of life for millions of people worldwide. Yet, conducting research in this area presents numerous challenges, including the risks of physical interaction between people and robots. Physics simulations have been used to optimize and train robots for physical assistance, but have typically focused on a single task. In this paper, we present Assistive Gym, an open source physics simulation framework for assistive robots that models multiple tasks. It includes six simulated environments in which a robotic manipulator can attempt to assist a person with activities of daily living (ADLs): itch scratching, drinking, feeding, body manipulation, dressing, and bathing. Assistive Gym models a person's physical capabilities and preferences for assistance, which are used to provide a reward function. We present baseline policies trained using reinforcement learning for four different commercial robots in the six environments. We demonstrate that modeling human motion results in better assistance and we compare the performance of different robots. Overall, we show that Assistive Gym is a promising tool for assistive robotics research.
I. INTRODUCTION
Assistive Gym is an open-source physics simulation framework for physical human-robot interaction and assistance. It models six activities of daily living, human preferences, and robot baselines, and reports improved task success when human movement is modeled.
I. INTRODUCTION
- Assistive Gym is an open-source physics-based framework for physical human-robot interaction, with environments for six activities of daily living.Tasks include itch scratching, drinking, feeding, body manipulation, dressing, and bed bathing.Source: §I · pp. 1–2
- Assistive Gym can benchmark control algorithms, support development of environments and controllers for new assistive tasks, and help compare robots for a task.The simulated tasks are described as commonly requested among adults who require physical assistance.Source: §I · p. 1
- The framework models human preferences and supports robots learning to prioritize assistance consistent with those preferences.The introduction connects preference-aware assistance with consistent, high-quality care.Source: §I · pp. 1–2
- 30.4% average task-success increase came from modeling human movement through co-optimization, with the human and robot trained simultaneously to achieve the same task.The authors frame this for cases where a person has limited motor functionality.Source: §I · p. 1
- The authors provide and analyze baseline policies for four commercial robots across the assistive tasks, enabling robot and controller comparisons.The robots are PR2, Sawyer, Baxter, and Jaco.Source: §I · pp. 1–2
A. Simulation Environments
Assistive Gym builds on simulation frameworks to focus on human-centric physical assistance. It provides common simulated tasks and a framework intended to support assistive robotics research.
A. Simulation Environments
- Assistive Gym builds on OpenAI Gym’s benchmark, common-interface, and comparison-tool model, while providing human-centric environments for direct assistance.Unlike manipulation-focused simulation frameworks, it centers tasks in which robots help people.Source: A. Simulation Environments · p. 2
B. Robotic Assistance
- Assistive Gym incorporates a co-optimization approach previously presented for a KUKA IIWA and person, extending it to multiple robots and assistive tasks.The earlier work jointly trained a robot and person for an assistive dressing task in simulation.Source: B. Robotic Assistance · p. 2
- The framework aims to support larger-scale safe learning, provide a common training framework and baselines, and establish canonical problems for assistive robotics.It is also intended to lower the barrier to algorithm development and support comparisons among control algorithms and robots.Source: B. Robotic Assistance · p. 2
III. ASSISTIVE GYM
Assistive Gym provides customizable physics-based environments for robots that physically interact with people. Its human models include configurable physical characteristics and limitations, alongside support for four commercial robots.
III. ASSISTIVE GYM
- Assistive Gym uses PyBullet and the OpenAI Gym interface to provide customizable physics-based environments for physical human-robot interaction and existing control-learning algorithms.PyBullet supports real-time CPU and GPU simulation, soft bodies and cloth, and programmatically created robot and human models.Source: §III · p. 2
A. Human Model and Robots
- The framework supports the PR2, Jaco, Baxter, and Sawyer, and includes default male and female human models with published 50th-percentile body sizes, weights, and joint limits.The human model has 40 controllable joints and can be modified programmatically.Source: A. Human Model and Robots · p. 2
- Human limitations include tremor, joint weakness, and limited range of motion, with one limitation randomly applied when a human model is created.The limitation parameters are sampled from specified ranges, and the simulator randomly selects the male or female model.Source: A. Human Model and Robots · pp. 2–3
B. Assistive Tasks (Environments)
Assistive Gym includes six simulated activities of daily living, each with task-specific tools, goals, and rewards. The tasks cover personal care, mobility, and feeding or drinking assistance.
B. Assistive Tasks (Environments)
- The six environments cover itch scratching, bed bathing, drinking water, feeding, dressing, and arm manipulation, with rewards tied to task-specific goals.The tasks use tools such as a scratching tool, washcloth, cup, spoon, gown, and scooping tool.Source: B. Assistive Tasks (Environments) · p. 3
- In itch scratching, the robot targets a random location on the right arm, applies less than 10 N near the target, and makes small scratching motions.Bed bathing rewards wiping the bottom of a washcloth along the right arm to clean distributed markers.Source: B. Assistive Tasks (Environments) · p. 3
- Drinking and feeding tasks bring a cup or spoon to the person’s mouth, while dressing pulls a gown sleeve along the left arm toward the shoulder.Drinking rewards pouring water into the mouth; feeding rewards placing food there without spilling it.Source: B. Assistive Tasks (Environments) · p. 3
- Arm manipulation uses a scooping tool to lift the person’s right arm from beside the bed and place it next to the torso.For most tasks, the robot holds a task-relevant tool, and its starting end-effector position is perturbed by up to 5 cm along each global axis.Source: B. Assistive Tasks (Environments) · p. 3
C. Realistic Human Joint Limits
Assistive Gym models pose-dependent human arm limits and selects robot base poses using task-centric kinematic criteria. These features aim to represent more realistic human movement and support task performance despite variation.
C. Realistic Human Joint Limits
- Fig. 2 contrasts a pose attainable with pose-independent limits with the more realistic pose reached when pose-dependent limits are enabled.The model makes the arm return to its previous valid configuration when a simulated pose is invalid.Source: Figure 2 · p. 3, C. Realistic Human Joint Limits · p. 3
- Pose-dependent limits constrain a joint’s range according to other joint configurations, and Assistive Gym incorporates a neural-network model trained on human motion-capture data to represent valid arm poses.At each time step, invalid right- or left-arm poses are reset to the preceding valid configuration.Source: C. Realistic Human Joint Limits · p. 3
D. Robot Base Positioning
- The baseline base-positioning method samples 100 robot base poses and selects one with collision-free inverse-kinematics solutions for the most goal end-effector poses.Ties are resolved by choosing the pose with the highest summed joint-limit-weighted kinematic isotropy (JLWKI).Source: D. Robot Base Positioning · p. 3
- PR2, Baxter, and Sawyer base poses are optimized at trial start, while Jaco is mounted at fixed wheelchair or nightstand locations depending on the task setting.The nightstand pose is optimized for bed bathing and arm manipulation.Source: D. Robot Base Positioning · pp. 3–4
E. Modeling Human Preferences
Assistive Gym models care preferences as costs that contribute to the reward alongside task success. Its unified baseline preference model is intended to apply across assistive tasks.
E. Modeling Human Preferences
- A unified baseline preference model spans assistive tasks, addressing preferences such as avoiding large forces and favoring slow, interpretable actions.The preference components are activated by task and assigned weights.Source: E. Modeling Human Preferences · p. 4
- The reward combines a state-based human-preference reward with the robot’s task-success reward, and both robot and human policies use the same overall reward.The preference reward reflects how well the robot satisfies the person’s preferences in the current state.Source: E. Modeling Human Preferences · p. 4
- Preference costs include high end-effector velocity, force directed away from the assistance target, and forces above 10 N near the target.The framework defines separate penalty terms for these behaviors.Source: E. Modeling Human Preferences · p. 4
- Other costs cover spilling food or water, delivering it at high velocity, garment force on the body, and high pressure from large tools.A separate high-pressure term allows large forces distributed over a broad surface for tasks such as arm manipulation.Source: E. Modeling Human Preferences · p. 4
IV. POLICY LEARNING AND CONTROL
Assistive Gym trains assistance controllers with reinforcement learning and specifies robot and human actions, observations, and actuator limits for the simulation.
IV. POLICY LEARNING AND CONTROL
- Robot actions are joint-position changes for 7-DoF arms, while human actions change arm or head positions; robot actuator strength is limited to reduce the likelihood of high forces.Action dimensions are 7 for a single robot arm, 14 for dual arms, 10 for the human arm, and 4 for the head.Source: §IV · p. 4
- Robot observations include end-effector pose, arm joint positions, applied forces, and task-relevant human joint positions, chosen to reflect information available in real assistance scenarios.Examples of human measurements include wrist, elbow, shoulder, and head pose.Source: §IV · p. 4
- The framework uses PPO to learn assistance policies, training robot and active-human policies concurrently with a shared reward during co-optimization.The policy network has two fully connected hidden layers of 64 nodes with tanh activations.Source: §IV · p. 4
V. EVALUATION
The evaluation trains and tests baseline policies for four robots across six assistive tasks. Policies handled some tasks reasonably, while dressing and arm manipulation remained challenging and all robots had room to improve.
V. EVALUATION
- The evaluation trained 24 robot-task policies, selected the best of three random seeds for each, and assessed policies over 100 rollouts.Training used 36 concurrent simulation actors and 10 million time steps per policy; dressing training took about six days because it simulated dynamic cloth.Source: A. Robotic Assistance · p. 5, §V · p. 5
A. Robotic Assistance
- PPO policies performed reasonably on itch scratching, bed bathing, feeding, and drinking, but dressing and arm manipulation remained challenging for all four robots.Dressing failures involved difficulty pulling a gown up the arm; arm lifting often used a tool edge, applying high pressure and earning a large negative reward.Source: A. Robotic Assistance · p. 5
- The static-human comparison reports average reward for each robot and task, with task-success rates for the highest-reward robot.The policies were evaluated over 100 simulation rollouts per robot-task pair.Source: A. Robotic Assistance · p. 5, Table I · p. 6
- All four robots showed significant opportunities for improvement across the six tasks with a static human; the PR2’s shorter arm span made itch scratching harder.Itch scratching can require reaching around the person’s arm.Source: A. Robotic Assistance · pp. 5–6
B. Collaborative Assistance
The authors model collaborative assistance by training active-human and robot policies together. Across almost all cases, robots achieved higher reward, and the best-performing robot averaged a 30.4% task-success increase over static-human assistance.
B. Collaborative Assistance
- Co-optimization trains separate PPO policies for the robot and human simultaneously, using a shared reward but different observations and actions.The human policy observes human joint angles, while the robot policy observes robot joint angles.Source: B. Collaborative Assistance · p. 6
- The active-human setup gives the person’s arm and head enough motor strength to lift against gravity, and evaluates trained policies over 100 rollouts.The evaluation retrains policies for all robots and tasks, selecting the best policy from three random seeds.Source: B. Collaborative Assistance · p. 6
- The best-performing robot achieved a 30.4% increase in task success on average when assisting an active rather than static human.The active-human evaluation excludes arm manipulation because a person can often lift their arm back onto the bed unaided.Source: B. Collaborative Assistance · p. 6, Table II · p. 6
VI. CONCLUSION
Assistive Gym is an open-source physics simulation framework focused on physical robot–human interaction. It models human capabilities and care preferences, and supports baseline policies, benchmarking, and robot comparisons across six assistive tasks.
VI. CONCLUSION
- The framework models human physical capabilities and preferences for assistance in physical robot–human interaction.The conclusion presents it as an open-source physics simulation framework for assistive robotics.Source: §VI · p. 6
- Baseline policies cover four robots and six assistive tasks, providing a basis for benchmarking and comparing robots.The framework is also demonstrated for developing environments for multiple assistive tasks.Source: §VI · p. 6
- Assistive Gym is presented as a promising open-source framework for developing autonomous robots that provide versatile physical assistance.Source: §VI · p. 6